In development

Code review for AI-written code

Review only what matters, in an order that makes sense.

humancr is a desktop app for reviewing code that AI agents write. It rates every changed file by how much it matters, so you spend your time where a mistake would hurt and skip the rest with confidence. It puts the files in an order you can follow, so you understand the change as you scroll instead of piecing it together at the end.

In development. humancr is not available yet. Windows, macOS and Linux

The humancr Changes view for a 15-file review. On the left, an importance filter with counts for High, Medium, Low, Safe and Unclassified files, and a review path that groups the files by feature. On the right, a syntax-highlighted diff of a C# change.
The Changes view: files rated by importance and read along a review path grouped by feature.

Built for the amount of code AI agents write

humancr turns an agent’s branch into a focused, merge-request-style review on your machine: the pull request review you know, with triage, a reading order and a direct line back to the agent.

Importance

Know where to look

Every changed file is rated Safe, Low, Medium or High. Safe files stay hidden until you want them, and unsure ratings round up. Your override survives re-analysis.

Review path

A path through the change

Read the files layered, foundations first, or clustered by feature in named groups. Drag a file to reorder, and your order sticks.

Agents over MCP

Closes the loop with your agent

A local MCP server lets your coding agent open reviews, explain non-obvious changes up front and pick up your feedback. Works with Claude Code and other MCP clients.

Smooth on huge diffs

The diff is never loaded as a whole: files and lines are virtualized, highlighting runs in background workers, and large or generated files, lockfiles and binaries start collapsed.

changed files in the benchmark review
5,416
changed lines across those files
393,733
from opening that review to the first diff rows on screen
0.44 seconds

Measured in the full app on a generated repository, on a 4-vCPU machine without a GPU.

How a review goes

Your agent and humancr talk over MCP. After you submit, humancr gives you a short message to hand to your agent.

  1. Your agent opens the review

    When its work is committed, the agent calls create_review with a description, then add_explanation for anything non-obvious.

  2. You review it

    Comment on a line, a range, a file or the whole review, with Markdown and suggestions. Comments stay drafts until you submit: Approve, Request changes or Comment.

  3. The agent follows up

    It fetches your review with get_review_feedback, fixes, replies, resolves and commits. New commits become a new version; “Since my last review” shows what changed.

A new database migration file in a humancr diff, rated High. Under lines 9 and 10, an explanation from claude-code says: Both columns are nullable so existing orders keep working without a backfill.
An agent’s explanation, attached to the lines it explains.

Runs on your machine

A desktop app that works with the tools you already have.

Your git, untouched

humancr uses your system git, so your credentials, config and hooks apply. It never merges, rebases, stashes or resets. Repositories in WSL work too.

Local reviews

Reviews are saved in a local database on your computer. The MCP server listens on 127.0.0.1 only and requires a token.

AI if you want it

Bring your own OpenRouter key, kept in your OS’s secure storage. Without one, everything else works, and your agent can rate files or set the order when you ask.

Secrets stay out

Keys, certificates and .env files are never sent to a model, secrets in diffs are redacted, and AI can be turned off per repository.

Status

In development

humancr is being built and isn’t available yet: there is nothing to download or sign up for. When it’s ready, you’ll find it here.